Inside the loop: how an AI investigator decides what to look at next

Most people-search tools work like a vending machine. You put in a name, they look it up in a database they built last year, and they hand back whatever matched. If the match was wrong, you find out later.

Nuculair works more like an analyst with a notebook. It doesn't look anything up in advance. It investigates live, one step at a time, and after every step it decides what's worth looking at next. That decision is the interesting part, so here's how it's made.

#It starts by breaking the case apart

A case starts from whatever you have: a name, an email address, a username, a phone number or a photo. The first thing the engine does is perceive: it reads that input, separates what you already know for certain from what you're asking, and splits the case into somewhere between three and eight threads.

A thread is a line of inquiry. For an email address, one thread might be "which services is this registered with", another "has it appeared in a breach", another "does the name in it match a handle elsewhere". Threads are what the engine spends its time on.

#Then it runs a loop, up to twenty times

Every cycle of the loop has the same steps. You can watch it run on the How it works section of our homepage.

#Select

Pick the thread with the most to gain. A thread that's one step from confirming an identity is worth more than one that's already been exhausted.

#Plan

Choose only the tools that can use this thread's input. Inputs are typed, so a username never gets sent to a tool that expects a phone number. That keeps collection fast and avoids wasting queries.

#Act

Run the collection, inside the privacy tier you set for the case. Passive cases only read public sources. Strict mode goes further: nothing identifying leaves your environment, and lookups use hashes.

#Shape

Turn raw results into typed entities: an account, a domain, a location, an organisation. Raw pages are kept as evidence; entities are what the engine reasons about.

#Interpret

Read the new entities through a set of specialist lenses, each asking a different question. Is this the same person? Is this infrastructure shared? Do the timelines line up?

#Challenge

This is the step most systems skip. Before anything is accepted, the engine argues against it. Is there a namesake who fits just as well? Is the account older than the person's first known handle? Is it in a different country? A link is only accepted once its confidence clears 0.8, and it's thrown out below 0.1. Everything in between stays open.

#Integrate

What survives goes into the case graph with its provenance attached: the tool that found it, the raw reference, when it was captured, its confidence and the privacy tier it was collected under.

#Why a loop and not a single search

Because the best next question depends on the last answer. Finding a username changes which tools are worth running. Finding a photo opens a reverse image search that didn't make sense before. A single search can't do that; a loop can.

The loop stops when twenty cycles are done or when no thread is worth another step. Most of the work in an investigation is deciding what not to follow, and the loop makes that decision explicit every cycle.

#The rule on top of the loop

One rule overrides everything above: the engine only expands what an operator confirms. A new photo match or a newly linked account can take a case somewhere new, and that's exactly where an automated system can latch onto the wrong person. So those pivots wait for a human. We explain why in Only expand what an operator confirms.

#What we'll publish next

A description of a method is only half the story. The other half is how often it's wrong. We're preparing a scorecard measured on synthetic personas, including same-name decoys built to trip the engine up, and we'll publish it with the misses included.